Reducing exercise-related hypoglycemia in automated insulin delivery with reinforcement learning

Abstract

Exercise is an important component for glucose management in type 1 diabetes, but remains challenging for automated insulin delivery systems as altered glucose dynamics are difficult to model explicitly. Glucose monitoring data might enable data-driven approaches for learning these dynamics implicitly. We propose combining model predictive control with a reinforcement learning component to adjust basal insulin infusion rates for exercise. We train our model on a variety of exercise scenarios and demonstrate improved glucose control using two different frameworks. We evaluate how generalizable both frameworks are by personalizing a trained model with a small number of additional individual-specific training episodes.